fdo-mirrors/mesa
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/mesa/mesa.git synced 2026-05-06 05:08:08 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

intel/brw: Make ifind_msb SSA friendly

Browse source 
No shader-db changes on any Intel platform.

v2: Use negate(tmp) instead of creating a new temporary. Suggested by
Ken.

fossil-db:

Meteor Lake, DG2, and Skylake had similar results. (Meteor Lake shown)
Totals:
Instrs: 152535897 -> 152535883 (-0.00%); split: -0.00%, +0.00%
Cycle count: 17112329592 -> 17112406110 (+0.00%); split: -0.06%, +0.06%

Totals from 40 (0.01% of 633223) affected shaders:
Instrs: 458813 -> 458799 (-0.00%); split: -0.01%, +0.00%
Cycle count: 4358016282 -> 4358092800 (+0.00%); split: -0.23%, +0.24%

Tiger Lake and Ice Lake had similar results. (Tiger Lake shown)
Totals:
Instrs: 150560511 -> 150560465 (-0.00%); split: -0.00%, +0.00%
Cycle count: 15484534441 -> 15482372893 (-0.01%); split: -0.12%, +0.11%
Spill count: 59795 -> 59794 (-0.00%)
Fill count: 103513 -> 103509 (-0.00%)

Totals from 40 (0.01% of 632445) affected shaders:
Instrs: 368877 -> 368831 (-0.01%); split: -0.01%, +0.00%
Cycle count: 3918398264 -> 3916236716 (-0.06%); split: -0.49%, +0.43%
Spill count: 16896 -> 16895 (-0.01%)
Fill count: 27819 -> 27815 (-0.01%)

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/30650>
This commit is contained in:
Ian Romanick 2024-08-09 18:25:36 -07:00 • committed by Marge Bot
parent e9c151fde6
commit c8038643b8
1 changed files with 8 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

13
src/intel/compiler/brw_fs_nir.cpp
View file

@ -1571,18 +1571,21 @@ fs_nir_emit_alu(nir_to_brw_state &ntb, nir_alu_instr *instr,
assert(instr->def.bit_size == 32);
assert(nir_src_bit_size(instr->src[0].src) == 32);
bld.FBH(retype(result, BRW_TYPE_UD), op[0]);
brw_reg tmp = bld.FBH(retype(op[0], BRW_TYPE_D));
/* FBH counts from the MSB side, while GLSL's findMSB() wants the count
* from the LSB side. If FBH didn't return an error (0xFFFFFFFF), then
* subtract the result from 31 to convert the MSB count into an LSB
* count.
*/
bld.CMP(bld.null_reg_d(), result, brw_imm_d(-1), BRW_CONDITIONAL_NZ);
brw_reg count_from_lsb = bld.ADD(negate(tmp), brw_imm_w(31));
inst = bld.ADD(result, result, brw_imm_d(31));
inst->predicate = BRW_PREDICATE_NORMAL;
inst->src[0].negate = true;
/* The high word of the FBH result will be 0xffff or 0x0000. After
* calculating 31 - fbh, we can obtain the correct result for
* ifind_msb(0) by ORing the (sign extended) upper word of the
* intermediate result.
*/
bld.OR(result, count_from_lsb, subscript(tmp, BRW_TYPE_W, 1));
break;
}